Scribd
Upload
163 views3 pages

The CUSUM Test: When The Regression Is Estimated Using Only The First T 1

The CUSUM test is used to detect structural changes in time-series data without specifying when the change occurred. It examines recursive residuals from a regression model over time to see if their distribution is changing, which could indicate instability. The test statistic is the cumulative sum of the scaled residuals, plotted against the observation number with confidence boundaries. If the plot strays outside the boundaries, it provides evidence against model stability. However, the CUSUM test has limited power compared to the Chow test, which requires knowing when the structural change took place.

Uploaded by

KAUSHIK BHATTACHARJEE
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
163 views3 pages

The CUSUM Test: When The Regression Is Estimated Using Only The First T 1

The CUSUM test is used to detect structural changes in time-series data without specifying when the change occurred. It examines recursive residuals from a regression model over time to see if their distribution is changing, which could indicate instability. The test statistic is the cumulative sum of the scaled residuals, plotted against the observation number with confidence boundaries. If the plot strays outside the boundaries, it provides evidence against model stability. However, the CUSUM test has limited power compared to the Chow test, which requires knowing when the structural change took place.

Uploaded by

KAUSHIK BHATTACHARJEE
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 3

The CUSUM Test

The technique is suitable for time-series data and may be employed if one is in doubt about

when a structural change could have taken place. The test is fairly general in nature in that it

does not need a beforehand specification of when the structural change actually occurs.

“Suppose that the sample contains a total of T observations The t-th recursive residual is the ex

post prediction error for yt when the regression is estimated using only the first t − 1

observations. Since it is computed for the next observation beyond the sample period, it is also

labeled a one step ahead prediction error”;

et  yt  xtbt 1

where xt is the vector of regressors associated with observation yt and bt−1 is the least

squares coefficients computed using the first t − 1 observations.

The forecast variance of this residual is

 2ft   2 [1  xt( X t1 X t1 ) 1 xt ]

Let the rth scaled residual be

er
wr 
[1  xr ( X r1 X r 1 ) 1 xr ]1/2

Under the hypothesis that the coefficients remain constant during the full sample period,

wr~ N[0, σ2] and is independent of ws for all s≠r . Evidence that the distribution of wr is changing

over time weighs against the hypothesis of model stability.


^

“One way to examine the residuals for evidence of instability is to plot wr /  simply against the

date. Under the null hypothesis (i.e. β is same in every period) of the model, these residuals are

uncorrelated and are approximately normally distributed with mean zero and standard deviation

1. Evidence that these residuals persistently stray outside the error bounds−2 and +2 would

suggest model instability”. (Some authors and some computer packages plot er instead, in which

 
case the error bounds are 2 [1  xr ( X r 1 X r 1 ) xr ]
1/2
ˆ

The CUSUM test is based on the cumulated sum of the residuals:

r t
wr
Wt  
r  K 1 ˆ
r T r T
where ˆ 2  (T  K  1) 1  (w
r  K 1
r  w) 2 and w  (T  K ) 1 
r  K 1
wr

“Under the null hypothesis, Wt has a mean of zero and a variance approximately equal to the

number of residuals being summed (because each term has variance 1 and they are independent).

The test is performed by plotting Wt against t. Confidence bounds for the sum are obtained by

plotting the two lines that connect the points [K,±a(T − K)1/2]

and [T,±3a(T − K)1/2]. Values of a that correspond to various significance levels . Those

corresponding to 95 percent and 99 percent are 0.948 and 1.143, respectively. The hypothesis is

rejected if Wt strays outside the boundaries.”

“However the ‘CUSUM’ test is frequently criticized on this basis that the power of the test is

rather limited when compared with that of the ‘Chow test’. The Chow test, however, is based on

a rather definite piece of information, namely, when the structural change takes place. If this is

not known or must be estimated, then the advantage of the Chow test diminishes considerably.”
The CHOW Test

You might also like